DecentVLA/pi05_cubestack_fl_3client

pi0.5 CubeStack — FedAvg 3-client (global)

Federated global model: FedAvg over 3 color-pair clients, 30 rounds x 250 local steps, full fine-tune. Shared pooled-6-repo normalizer. The FL method point vs the centralized ceiling and the local-only floor.

lerobot-native export (portable tokenizer, real 6-D action/state; --policy.compile_model=false). SO-101 CubeStack, cameras base_0_rgb/left_wrist_0_rgb. Trained on Isambard-AI (GH200) with decent-vla. Part of the 3-client pi0.5 CubeStack federated study (color-pair non-IID: c0=harry Green/Orange, c1=zhekai Green/Blue, c2=kevin Orange/Blue; each client blind to the 3rd color).